Developing General Reactive Element-Based Machine Learning Potentials as the Main Computational Engine for Heterogeneous Catalysis

Machine learning potentials (MLPs) have emerged as a promising technique to significantly enhance efficiency by replacing computationally expensive quantum mechanical calculations. However, developing truly universal MLPs remains challenging, as the consensus is that MLPs can only be used for similar structures that they have been trained on, while the vast and diverse chemical space is difficult to fully sample using the common system-dependent sampling methods. Here, our approach leverages a unique random exploration via imaginary chemicals optimization (REICO) strategy, which enables unbiased exploration of chemical space by focusing on atomic interactions. The resulting EMLP is inherently general and reactive, capable of accurately predicting elementary reactions without explicit structural or reaction pathway inputs. Benchmarked across various representative calculations of heterogeneous catalysis, our EMLP achieves quantitative agreement with density functional theory (DFT) calculations. This demonstrates the potential of EMLP as a powerful, general, and user-friendly tool for modeling complex chemical systems, paving the way to replace DFT calculations for large and intricate systems. Our approach is also applicable to broader fields such as materials science and molecular biology, representing a paradigm shift in MLPs-related research.